Unscented Kalman Filtering for Articulated Human Tracking

We present an articulated tracking system working with data from a single narrow baseline stereo camera. The use of stereo data allows for some depth disambiguation, a common issue in articulated tracking, which in turn yields likelihoods that are practically unimodal. While current state-of-the-art trackers utilize particle filters, our unimodal likelihood model allows us to use an unscented Kalman filter. This robust and efficient filter allows us to improve the quality of the tracker while using substantially fewer likelihood evaluations. The system is compared to one based on a particle filter with superior results. Tracking quality is measured by comparing with ground truth data from a marker-based motion capture system.